Why chain prompts
One large prompt can hide errors and make revisions expensive. A chain separates research, planning, drafting, and review so each intermediate result can be checked.
Design the handoff
Give every step an input contract and output schema. Pass only information the next step needs and retain source identifiers when traceability matters.
Add gates
Validate structure, required evidence, and business rules between steps. Route failures to a repair step instead of blindly continuing.
Control cost and drift
Cache safe intermediate results, cap retries, version each prompt, and test the full chain. Optimizing one step can still reduce end-to-end quality.
Worked example
Step 1: Extract claims from [DOCUMENTS], each with a source ID.
Gate: Reject claims without a supporting passage.
Step 2: Group accepted claims into themes.
Step 3: Draft a summary using those themes only.
Gate: Check that every summary claim maps to an accepted claim.How to check the result
Introduce a malformed result at the first handoff. The chain should stop or repair that result rather than continue. Measure the final summary, not just the apparent quality of intermediate steps.